No description available.

No description available.

No description available.

No description available.

The goal of this dissertation is to introduce a method of fault diagnosis that is designed to cope with fault signature variability, the main source of difficulty for the existing diagnostic systems. This method is a non-parametric pattern classifier that uses a multi-valued influence matrix (MVIM) as its diagnostic model. In this method, process abnormalities are detected through processing the sensory data and flagging, and diagnostic reasoning is performed by matching the flagged measurements against the columns of the influence matrix. Fault signature improvement is achieved by a Flagging Unit, which is tuned based on a training set. This unit is shown to have the ability to improve detection, reduce false alarms, and enhance diagnostics. The improved fault signatures by the Flagging Unit are also shown to be beneficial to other classifiers such as the Bayes classifier and artificial neural nets. The applicability of the MVIM method is investigated in fault diagnosis of a helicopter gearbox. A total of five tests were performed, during which eight failures occurred. In order to enhance the effect of the failures on the vibration data, the vibration signals obtained from the gearbox were digitized and processed by a vibration signal analyzer. The parameters obtained from this signal analyzer were then utilized to train the MVIM method and test its performance for both detection and diagnosis. The averaged values of the parameters obtained from individual accelerometers were used to reduce the processing time. Training sets were formed based on parameters from various combinations of the five tests, and the MVIM method was tested based on the parameters from all of the five tests. Detection results indicate that the MVIM method provided excellent results when the full range of faults' effects on the vibration measurements were included in the training set. The MVIM method was also utilized to rank the parameters for their significance in detection. It is shown that through this ranking the optimal subset of parameters for detection can be selected, which is particularly important in reducing processing time for on-line detection. For diagnosis, the MVIM method was used in a hierarchical manner. The parameters from individual accelerometers were first processed through detection MVIMs, to trigger the presence of a fault, and then examined by diagnostic MVIMs to identify the fault. Diagnostic results show that the MVIM method had a correct diagnostic rate of 95% for the faults included in training. (Abstract shortened by UMI.)

The problem of detection of a constant signal in additive noise is addressed. In centralized detectors all observations are obtained and processed at the same place, whereas in distributed detectors observations on the same phenomenon is obtained at several geographically dispersed local sensors. The local sensors transmit a condensed signal to a global decision maker. In this dissertation we address some problems under both categories. We present centralized truncated sequential nonparametric detectors that can be implemented with a hard-limiter and a dead-zone limiter. These detectors are based on approximations of the sequential sign and the sequential conditional sign detectors. The sequential tests are modelled as Markov chains for design and performance evaluations. By using truncation the possibility of excessively long tests is removed. The approximations allow mathematically tractable design for constant false alarm rate (CFAR) performance. In comparison to the sequential sign and conditional sign detectors, the proposed detectors have slightly higher average sample numbers (ASN) under no signal and nominal signal conditions, but considerably lower ASN for intermediate signal strengths. A number of distributed detection schemes are considered. Firstly, the optimum decision policy for a sequential fusion center with fixed local sensors is studied. The sequential fusion rule is numerically studied using Markov chain modelling. We then focus on deriving optimum M level quantizers at the local sensors where $M > 2$. The solutions are obtained assuming a fixed fusion rule using the Bayesian and the locally optimum detection criteria. The receiver operating characteristics for different values of M and a number of different fusion rules are compared. The performance improves with increasing values of M; however the communication cost also increases. Finally, the concept of multilevel quantization is extended to distributed nonparametric detection. Two schemes are presented which employ Wilcoxon statistics and generate M-level signals from the local sensors. Design issues and numerical performance evaluations of the proposed detectors are presented.

No description available.

No description available.

No description available.

No description available.